Credit Card Customer Segmentation¶

We are working for a credit card company and are given a dataset that contains information about the company’s clients. We're asked to help segment them into different groups in order to apply different business strategies for each type of customer. Possible strategies are:

• Provide higher credit limits for customers that use the card a lot, but spend little money.
• Create incentives for those with high income who don't use the card as much as the company expects.

The company expects to receive a group for each client and an explanation of the characteristics of each group and the main points that make them different.

Data dictionary:

Outline¶

• Exploratory Data Analysis
• Prepare the data for modeling
• Feature Selection
• Selecting Number of Clusters
• Interpreting Results
• Feature Engineering - Create interaction features
  • Feature Selection
  • Selecting Number of Clusters
  • Interpreting Results
• Conclusion

Exploratory Data Analysis¶

Back to the top

EDA observations:

1. There are no missing values. Dataset has 14 features and 10,127 observations. 3 of these features are categorical: ['gender', 'education_level', 'marital_status']
2. There is a strong correlation between months_on_book and age features
3. total_trans_count and total_trans_amount are also highly correlated
4. avg_utilization_ration and credit_limit have a non-linear relationship
5. Dataset has roughly balanced proportion of the two genders.

Prepare the data for modeling¶

• customer_id: this is the unique identifier for each customer and shouldn't be used for machine learning. we need to drop this.
• categorical columns need to be encoded.
  • The gender column contains only 2 unique values, "M" and "F", which means it's possible to replace all the rows with 1 and 0
  • The education_level column contains 6 unique values. This column represents levels of education, and it's possible to rank the levels from the lowest to the highest. Ordinal encoding is particularly beneficial in clustering algorithms like K-means, which rely on distance calculations. By assigning numerical values that reflect the natural progression of education levels, the algorithm can better capture the underlying structure in the data, avoiding the potential pitfalls of treating ordered categories as unrelated. However, it is important to ensure that the ordinal mapping accurately reflects the real-world progression and that the resulting numerical values do not distort the distance metrics in the clustering process.
  • The marital_status column has 4 unique values. We will use one-hot-encoding to create dummy variables for this column.
• All features need to be scaled, not just numeric columns.
  • For clustering, StandardScaler() is generally better than MinMaxScaler(). StandardScaler centers data around zero with unit variance, making it less sensitive to outliers—critical for distance-based algorithms like K-means.
  • Even though one-hot encoded columns are often left unscaled for tree based classification models, but clustering algorithms like K-Means are fundamentally different — it's a distance-based algorithm. It computes Euclidean distances between data points. Scaling ensures that all features contribute equally to the clustering process.

Back to the top

Feature Selection¶

There is a strong correlation between following set of features:

• months_on_book and age
• total_trans_count and total_trans_amount

While clustering algorithms like k-means are not fundamentally broken by correlated features, including highly correlated variables can disproportionately influence distance calculations, effectively giving them more weight in cluster formation. This redundancy can lead to biased results and increased computational overhead, especially in high-dimensional datasets.

So in order to avoid any biased results, we will drop one of the features from each set. Before we do that, let's visualize hierarchical clustering on a correlation matrix using dendrogram.

Back to the top

Dendrogram confirms our previous observation. So we will select one representative from each cluster based on domain relevance.

Given the business strategies:

1. Provide higher credit limits for customers that use the card a lot, but spend little money
2. Create incentives for those with high income who don't use the card as much as expected

The more domain-relevant feature between age and months_on_book is months_on_book. Here is why:

• months_on_book reflects customer tenure — how long a customer has been with the bank. This directly relates to behavioral history and engagement, which are critical for identifying underutilized high-income customers or frequent low-spenders.
• While age is correlated with months_on_book (often younger customers are newer), tenure is more actionable for segmentation because it captures relationship length and loyalty, which aligns better with credit limit decisions and incentive targeting.
• For strategy #2, identifying high-income customers who don’t use the card much benefits from knowing how long they’ve been customers — a long-tenured but inactive high-income customer may warrant different incentives than a new one.

Between total_trans_amount and total_trans_count, the more relevant is total_trans_count. Here is why:

• total_trans_count captures how frequently a customer uses the card — directly aligning with usage behavior.
• Strategy #1 targets high-frequency, low-spend users — we need transaction count to identify high frequency users
• Strategy #2 focuses on low usage among high-income customers — again, need total_trans_count refers to find low frequency users

So let's drop 'age' and 'total_trans_amount'.

Now our dataset is ready for machine learning.The next step is to decide the best number of clusters to use in the segmentation.

Selecting Number of Clusters: Elbow Curve and Silhouette analysis¶

We will use Elbow Curve and Silhouette analysis for this. We will use K-Means algorithm to create clusters. Elbow curve is a good starting point but it is important to keep business strategies in mind and that will help us with this decision. Here are the strategies of our company:

• High usage vs. low spending → Need to look for clusters with high total_trans_count but low avg_utilization_ratio. These customers demonstrate responsible credit behavior (low default risk) and high engagement, making them ideal candidates for higher credit limits to encourage more spending and increase issuer revenue through interchange fees, without significantly increasing risk.
• High-income, low engagement → Need to identify clusters with high estimated_income, low total_trans_count, and high months_inactive_12_mon.

We will also look into:

• credit_limit and avg_utilization_ratio to assess underutilized potential.
• months_on_book to distinguish loyal vs. new customers.
• total_relationship_count to identify deep-engagement opportunities.

Back to the top

The best possible candidate here is five clusters. Let's look at that options and interpret results.

Interpreting Results - Numeric Variables¶

Now, we will interpret the results and summarize the characteristics of each cluster and differentiate them from each other based on the variables used for the segmentation.

First, we'll analyze the numerical variables and see how they behave in each cluster.

Back to the top

We can see that cluster 0 is largest while cluster 5 is the smallest.

Considering the numeric variable only, we'll check on the average value of each variable per cluster. We just need to group the data and plot a bar chart for each column.

For those numerical variables with higher correlations we saw earlier, we can also use a scatter plot to visualize this correlation grouped by clusters and analyze how the clusters change between each area of the chart.

We can draw some early conclusions considering only the numeric variables.

For instance, Cluster 1 has the highest amount of money transitioned, while Cluster 4 has the highest income.

For the categorical columns, we'll plot the percentual distribution of each variable in each cluster. This will allow us to verify if a particular cluster is mostly composed of men, or of married people only, for example.

Considering the categorical variables, we notice that the education level is well divided between clusters. Cluster 0 is composed almost entirely of married people, while we don't know the marital status of anybody in Cluster 3. Cluster 4 is almost completely male and Cluster 2 is 100% made of single people.

Let's list the most important characteristics of each cluster and also make some suggestions and share our insights for each one of them.

Cluster 0¶

Characteristics: Mostly women; mostly married; low estimated income; low credit limit; low amount transitioned; high utilization rate.

Insight: Married people (majority women) with low income and limit but utilize too much of their credit with a few larger purchases

Cluster 1¶

Characteristics: Mostly men; mostly single and married, high credit limit; high amount transitioned; high number of transactions; low utilization rate; low total relationship count

Insight: People with high volume spent on the card, but low utilization. Could be incentivized to spend more and targeted for more engagement.

Cluster 2¶

Characteristics: Mostly women; 100% single people, low estimated income, low credit limit, low amount transitioned, high utilization rate.

Insight: Single (mostly women) people that use their card a lot but have low credit limits and income. Could be given a bit more credit limit.

Cluster 3¶

Characteristics: Gender well divided; low credit limit, high utilization rate; marital status 100% unknown; smaller cluster.

Insight: Men and women with low credit limits but do have high balances.

Cluster 4¶

Characteristics: Mostly men, mostly single and married, high estimated income, high credit limit; low amount transitioned; low utilization rate.

Insight: People (majority men) with high income and credit limits, but don't use the card. Could be incentivized to use it.

Cluster 5¶

Characteristics: Gender well divided, 100% divorced, low credit limit, low amount transitioned, high utilization rate.

Insight: Low credit limit and transactions, but use the card very often. Could receive benefits to spend more money.

Feature Engineering: Create interaction features¶

Let's explore adding iteraction featues to see how it works.

Back to the top

Feature Selection - using interaction features¶

Back to the top

• High usage vs. low spending → Need to look for clusters with high transactor_score, trans_per_balance, high trans_count_to_limit
• High-income, low engagement → Need to identify clusters with high income_inactive_score and high income_to_utilization

Drop highly correlated features¶

Selecting number of clusters - using interaction features¶

Back to the top

Interpreting Results - using interaction features¶

Back to the top

We need to apply log transformation only to skewed numerical features like estimated_income or credit_limit. We will use log(1 + x) to handle zero values (np.log1p does this).

• High usage vs. low spending → Need to look for clusters with high transactor_score, high trans_per_balance and high trans_count_to_limit
• High-income, low engagement → Need to identify clusters with high income_inactive_score and high income_to_utilization

transactor_score (Strategy 1) and high_income_low_use (Strategy 2) are clustering together in cluster 2. This indicates overlapping behavioral patterns.

Interpreting results - categorical columns¶

We can observe from the bar plots above:

• education_level: Cluster 2 has similar proportions as other clusters.
• marital_status: Cluster 0 and 1 has a single class but Cluster 2 has all 4 classes.
• gender: Cluster 2 has higher number of females than males, while other clusters are roughly balanced.

Conclusion¶

We first tried clustering without any engineered features. With this method we were able to find clusters with following characteristics:

Cluster 0¶

Characteristics: Mostly women; mostly married; low estimated income; low credit limit; low amount transitioned; high utilization rate. Insight: Married people (majority women) with low income and limit but utilize too much of their credit with a few larger purchases

Cluster 1¶

Characteristics: Mostly men; mostly single and married, high credit limit; high amount transitioned; high number of transactions; low utilization rate; low total relationship count Insight: People with high volume spent on the card, but low utilization. Could be incentivized to spend more and targeted for more engagement.

Cluster 2¶

Characteristics: Mostly women; 100% single people, low estimated income, low credit limit, low amount transitioned, high utilization rate. Insight: Single (mostly women) people that use their card a lot but have low credit limits and income. Could be given a bit more credit limit.

Cluster 3¶

Characteristics: Gender well divided; low credit limit, high utilization rate; marital status 100% unknown; smaller cluster. Insight: Men and women with low credit limits but do have high balances.

Cluster 4¶

Characteristics: Mostly men, mostly single and married, high estimated income, high credit limit; low amount transitioned; low utilization rate. Insight: People (majority men) with high income and credit limits, but don't use the card. Could be incentivized to use it.

Cluster 5¶

Characteristics: Gender well divided, 100% divorced, low credit limit, low amount transitioned, high utilization rate. Insight: Low credit limit and transactions, but use the card very often. Could receive benefits to spend more money.

Then we introduced interaction features into the dataset defined for each strategy. With this feature engineering, we have been able to segment customers for both strategies - High usage, low spending and High income low engagement. But both set of customers are clustering together. To further improve and to separate them, we can try following:

1. Use Gaussian Mixture Models (GMM) instead of KMeans — it allows soft assignments and better handles overlapping tendencies.
2. Cluster each strategy separately by filtering data and applying focused clustering.
3. Increase feature weighting for key differentiators (e.g., scale avg_utilization_ratio or months_inactive_12_mon) to force separation

Back to the top

Author¶

Puneet Pawar